Research On Deep Learning-based Intra Cross-Component Chroma Prediction And Luma Quality Enhancement For H.266/VVC

The continuous growth of video data has put forward higher requirements for video coding efficiency.At present,the latest generation of international video coding standard H.266/VVC has been released.Compared with the previous generation of international video coding standard H.265/HEVC,H.266/VVC still adopts a hybrid coding framework.By optimizing the algorithm of each module in the coding framework,it can achieve about 50%coding efficiency improvement.In recent years,due to the application of deep learning in the field of video coding,the efficiency of video coding has been further improved.This thesis aims to combine deep learning with the latest generation of international video coding standard H.266/VVC to further improve the efficiency of H.266/VVC.The main innovations are as follows:To improve the accuracy of intra chroma prediction,based on the analysis of the existing deep learning-based intra cross-component chroma prediction technology,we propose an intra cross-component chroma prediction method based on deep learning.The proposed neural network model is composed of the adjacent reference pixel feature extraction module,the current luma block feature extraction module,and the feature fusion module.The usage of the channel attention mechanism in the feature extraction block can explore the relationship between the adjacent reference pixels and the chroma components.Experimental results demonstrated that the proposed method can adapt different quantization parameter(QP)settings.By implementing the proposed method into the standard H.266/VVC test platform,average 2.89%and 2.36%bitrate savings can be achieved for Cb and Cr components,respectively,while preserving the same reconstruction quality.To improve the accuracy of intra luma prediction,a quality enhancement method for intra luma prediction based on deep learning is proposed.First,based on the analysis of the deep learning-based loop filter technology,a neural network model based on neuron attention mechanism is proposed,which consists of a shallow feature extraction module,a deep feature extraction module,and a reconstruction module.Among them,the shallow feature extraction block takes the normalized QP as input so that the network model can adapt to different QPs.The deep feature extraction module is composed of the proposed Multi-scale and Neuron Attention(MSNA)module in which a lightweight neuron attention mechanism is introduced.The number of MSNA modules is also explored through experiments to reduce the model complexity as well as ensuring the coding efficiency.Experimental results demonstrate that average 1.24%,0.34%,and 0.16%bitrate savings can be achieved for Luma,Cb and Cr components,respectively,while preserving the same reconstruction quality.